Amjad Masad & Adam D’Angelo: How Far Are We From AGI?

Summary

• Adam D’Angelo sees no LLM plateau: reasoning, code generation, and video improved so sharply in one year that computer-use objections and many current critiques should fade within another year or two. His practical AGI threshold is software better than a typical remote worker, and he expects “a very different world” five years out without requiring a fundamentally new architecture.
• Amjad Masad separates economically potent “functional AGI” from intelligence that can efficiently learn inside any new environment. Brute-force labeling, expert data, and contrived RL environments can automate many job components, but that dependence on “enormous effort and money and data” makes him doubt LLMs are naturally scaling toward true AGI or a singularity.
• Cheap digital labor could drive growth far beyond 4%-5%, but partial automation may first hollow out the human-capital pipeline. D’Angelo worries that experts could manage hundreds of agents while firms stop hiring entry-level workers; Masad adds that displacing experts could eventually undermine the data and expertise needed for further improvement.
• The scarce asset increasingly becomes whatever is missing from training data, creating demand for human knowledge, expert labeling, and verifiable RL environments. An AI might eventually prove every theorem from supplied axioms, D’Angelo argues, yet still cannot know “how did this particular company solve this problem 20 years ago?” unless someone records it.
• AI may simultaneously strengthen hyperscalers and unleash far more solo entrepreneurs, producing a barbell rather than a clean incumbent-versus-startup outcome. Competition among model suppliers keeps application costs falling, subscriptions monetize from day one, and weaker network effects permit multiple venture-scale winners—while founder-controlled incumbents are unusually capable of investing through disruption.
• Masad sees AI-assisted software creation as one of AI’s largest remaining unlocks, even though today’s vibe-coding tools remain far below professional engineering. His upside case is that anyone could eventually create what once required “a team of 100 professional software engineers”; he also wants more experimental research and less of Silicon Valley’s “get-rich driven” culture.
• Replit’s roadmap shows where agent economics are heading: longer autonomous loops, built-in verification, and parallel work across five to 10 agents before potentially reaching hundreds. Its progression ran from roughly two minutes for Agent V1 to 20 minutes for V2, an advertised 200 minutes for Agent 3, and actual user runs beyond 28 hours; Agent 4 is aimed at coordinated feature development and code merging.

Deep dive

1. Current LLM progress looks faster, not slower

• D’Angelo’s opening evidence against bearishness is the change within a single year: reasoning models, code generation, and video generation all improved rapidly. “I honestly don’t know what people are talking about,” he says; the apparent disappointment may reflect expectations that entire jobs would already be automated end to end.

• His practical definition of AGI is deliberately economic: a system able to outperform a typical remote worker at that worker’s job. Whether it also beats the world’s best individual or whole teams belongs to the AGI-versus-ASI terminology debate; the remote-worker threshold is the point where society becomes “a very different world.”

• The immediate constraint, in D’Angelo’s view, is often not intelligence but supplying the right context. Computer use remains immature, yet he expects it to work within “the next year or two,” unlocking automation of a large portion of human activity even if the resulting system does not satisfy everyone’s AGI definition.

• Memory and continuous learning are awkward for current architectures, but D’Angelo thinks they might be approximated well enough. Reasoning progress is “incredible,” pretraining is still improving, and he sees neither a demonstrated limit nor a need for an entirely new architecture.

2. Functional AGI can arrive without cracking intelligence

• Masad’s objection is partly political: predictions such as AI 2027 and the situational-awareness papers strike him as “not really science, they’re just vibes.” Claims that the whole economy will soon be automated risk frightening policymakers into bad regulation while outrunning the progress actually visible.

• His compact demonstration of LLM unevenness is counting letters in a sentence: three of four models failed, while GPT-5 with high thinking needed roughly 15 seconds. These are “amazing machines,” he says, but a different kind of intelligence whose limitations are being papered over through training data and surrounding infrastructure.

• Masad calls the emerging regime “functional AGI”: companies can collect expert data and construct RL environments for coding, investment banking, and other occupations until many functions become automatable. The cost is extensive labeling, contracting, human expertise, compute, and bespoke environment design—not the more direct scaling associated with GPT-2, GPT-3, GPT-3.5, and perhaps GPT-4.

• His preferred AGI test comes from older reinforcement-learning thinking: place a machine in any environment and let it learn efficiently, as a person could learn basic pool in two hours. D’Angelo concedes humans learn new skills from less data, but argues brute force can still create software as capable as an average worker—and economic consequences depend on capability, not elegance.

3. Automation may outrun the labor pipeline

• D’Angelo’s macro thought experiment starts with digital labor that costs $1 an hour and can perform any human job. That world should produce growth “much more than four to 5%,” but the system might instead cover only 80% of tasks, cost more than humans, or remain constrained by energy, power plants, and supply chains. Reaching the full capability could take five, 10, or 15 years.

• D’Angelo worries about an intermediate equilibrium: AI handles entry-level quality assurance while very good QA professionals manage hundreds of agents and cover long-tail failures. Productivity rises, yet firms stop hiring beginners because agents outperform them—leaving fewer people able to accumulate the experience required for expert work.

• Erik points to the same pattern among computer-science graduates, and D’Angelo says LLMs are likely contributing. Masad adds a recursive problem: if models require expert labels but displace the workers who become experts, where does the next improvement cycle’s training signal come from?

• Perfect environments such as AlphaGo can drive systems beyond expert level, D’Angelo notes; most jobs provide much weaker feedback. Masad expects near-term demand to favor people who use AI to accomplish what AI cannot complete alone. In a hypothetical world where everything is automated, he imagines more people pursuing art, poetry, and hobbies if wealth is distributed; he doubts every job will be automated in the current paradigm and places the growth of AI-leveraging job categories in the at-least-10-years range.

4. Human taste may be defensible—but data can overwhelm intuition

• Masad doubts the current paradigm will automate every occupation because many jobs consist of serving other humans. Without embodiment in human experience, he argues, machines cannot fully understand what people want; humans therefore remain the economy’s generators of ideas.

• D’Angelo’s counterexample is recommendation: Facebook, Instagram, and Quora ranking systems are already “superhuman” at predicting what an individual will read. A person can simulate another person’s preferences, but cannot match an algorithm trained on that user’s extensive click data, everyone else’s click data, and the similarities among those datasets.

• The chef tasting a dish illustrates the discussion’s human-judgment example: creators test work through their own senses. D’Angelo’s reservation is that any one chef or artist has “very little data compared to what AI can be trained on,” leaving the eventual balance between embodied judgment and aggregate behavioral data unresolved.

• Torenberg’s Sovereign Individual scenario extends the argument into politics: highly leveraged entrepreneur-capitalists spin up agent-powered companies while large populations become potentially unemployed or economically noncontributing. States may then compete for productive individuals, even negotiating tax rates, because political structures built around every citizen being economically productive would no longer fit. He presents this as a prediction worth considering, not an endorsement.

5. AI creates a barbell of hyperscalers and entrepreneurs

• Masad is most excited by the “vastly increased” capacity of one person. Ideas once blocked by recruiting, fundraising, and assembling complementary skills can now be brought into existence by solo entrepreneurs; he calls the newly broad availability of entrepreneurial opportunity “the most exciting thing about this technology.”

• Torenberg asks whether AI empowers the center or the edges. D’Angelo says that if it empowers the edges, there may be a barbell: incumbents become “much, much bigger” while individuals flourish at the edge, rather than technology resolving neatly into centralization or decentralization.

• D’Angelo sees a healthy market balance: competition among hyperscalers and labs gives applications alternatives and drives prices down “incredibly quickly,” but suppliers such as Anthropic and OpenAI can still raise enough capital for long-term investment. He expects both hyperscaler growth and a large crop of new application companies.

6. Incumbents learned the disruption playbook, while network effects weakened

• Masad agrees with D’Angelo that AI is both sustaining and disruptive. ChatGPT initially counterpositioned against Google’s trusted-information business because hallucination looked dangerous; Google, despite possessing internal chatbot technology, waited roughly two years to release Gemini, then used AI to improve Search overviews, Workspace, phones, and the rest of its products.

• Torenberg’s amendment is that “everyone read the book.” Public-market investors now punish incumbents for failing to adapt, leadership teams anticipate disruption, and founder control makes it easier to absorb near-term pain. In a less competitive environment like the 1990s, he thinks AI might have displaced incumbents more cleanly.

• Torenberg notes that passing on a company because it cannot become the single category winner imported too much Web2 logic. Masad argues that network effects matter less today: scale still supplies data and capital, but no longer makes smaller competitors impossible, leaving room for multiple foundation-model and application winners.

• Monetization also starts earlier. Unlike advertising businesses that needed millions or tens of millions of users, subscription products can charge immediately, aided by infrastructure such as Stripe. Geopolitical fragmentation adds another dimension: Torenberg suggests that an “OpenAI of Europe” or separate Chinese ecosystem may matter even when a global leader already exists.

7. Poe and Quora are bets on model diversity and scarce human knowledge

• Poe began with Quora’s early-2022 experiments comparing GPT-3 answers with human answers. The model responses were worse, but instant private answers to almost anything revealed a distinct product opportunity; Poe was conceived as an addition to Quora rather than simply an act of self-disruption.

• It was also an early bet that models would diversify. That thesis now spans image, video, audio, reasoning systems, and agents, making a general aggregator more useful. Even nontechnical consumers switch among ChatGPT, Gemini, and Claude for different tasks or “personalities”—behavior Masad did not expect from the Google era.

• D’Angelo expects data to become the bottleneck as intelligence gets cheaper. Scale AI, Surge AI, Mercor, and a long tail of newer firms are developing around putting human knowledge into a form models can use, while Quora positions itself as a source of human knowledge and uses AI to improve moderation, answer ranking, and its product experience.

8. Replit is stretching agents from minutes into coordinated workforces

• Masad frames the progression from Copilot autocomplete to chat, Cursor’s larger-file editing, and Replit’s agent as an expansion of scope. Replit’s system edits code, provisions databases and infrastructure, performs migrations, deploys, executes tests, reads errors, and repeats the debugging loop across the development lifecycle.

• The beta arrived in September 2024; Agent V1 followed around December as models moved from Claude 3.5 to 3.7. V1 ran roughly two minutes, V2 about 20, and Agent 3 was advertised at 200 minutes—though users have run it beyond 28 hours and Masad says it can operate “kind of indefinitely.”

• The unlock was verification. Inspired by a DeepSeek paper from NVIDIA on writing CUDA kernels, Replit put testing and computer use inside the loop so the agent could inspect an application, diagnose failures, and rewrite code. Masad stresses that longer runtime is not itself the goal: the system still needs to become faster, cheaper, and more reliable.

• Agent 4’s major direction is parallelism: request a login page, Stripe checkout, and admin dashboard, then let the AI identify which tasks can run concurrently and merge changes across the code. Masad expects developers to manage five to 10 agents initially, perhaps hundreds eventually, followed by multimodal whiteboards, persistent project memory, specialized agents, and Slack-like AI coworkers.

9. Vibe coding is underhyped, but research needs more play

• Masad calls vibe coding “unbelievably high-potential” precisely because today’s tools remain far below professional engineering. If they reach that level in a few years, everyone could build software that previously demanded 100 engineers, opening opportunities well beyond conventional application development.

• Asked whether he would study computer science in 2025, D’Angelo says yes: algorithms and data structures help people understand what is possible and manage agents. Every alternative discipline also faces automation arguments, so his advice is to study what one enjoys.

• Masad wants more “mad-science experiments”: DeepSeek-OCR’s apparent use of text screenshots to economize context, text diffusion built from a BERT instance masking words, and new combinations of pretrained, RL reasoning, encoder-decoder, and diffusion components. Silicon Valley’s increasingly “get-rich-driven” culture, he argues, leaves too little room for playful, strange prototypes.

• Torenberg observes that Claude 4.5 appeared more aware of dwindling context, economized tokens near the limit, and better recognized red-team or test settings. Masad still calls consciousness fundamentally nonscientific and points to Roger Penrose’s argument that the brain is not simply a computer because humans can detect things that Turing machines cannot. As a student, he would pursue philosophy of mind and neuroscience.